A Sensitivity Analysis Method and System for Gradient Parameters of Nuclear Power Plant Main Steam System Based on KAN

By using a KAN-based gradient parameter sensitivity analysis method combined with a data-driven surrogate model, the problem of low calibration efficiency of nuclear power plant main steam system models was solved, achieving efficient and accurate parameter sensitivity analysis and improving the safety and operational stability of nuclear power plants.

CN123116801APending Publication Date: 2026-11-17CHINA NUCLEAR POWER OPERATION TECH CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611263407.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, the high-dimensionality and nonlinear characteristics of the model parameters of the main steam system of nuclear power plants lead to low model calibration efficiency and poor accuracy of local sensitivity analysis, while global sensitivity analysis has high computational cost, making it difficult to perform parameter sensitivity analysis efficiently in nuclear power plants.

Method used

We employ a KAN-based gradient parameter sensitivity analysis method, combined with a data-driven surrogate model. By constructing a KAN model, we perform undirected and directed sensitivity analyses to assess the importance of each model parameter. We then utilize the KAN-GSA framework for efficient analysis.

Benefits of technology

It improves the accuracy and efficiency of sensitivity analysis of parameters in nuclear power main steam systems, reduces computational costs, can quickly identify the magnitude and direction of the influence of key parameters, improves the calibration accuracy and efficiency of the model, and is suitable for real-time monitoring and optimization under complex operating conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116801A_ABST
    Figure CN123116801A_ABST
Patent Text Reader

Abstract

This invention relates to a KAN-based gradient parameter sensitivity analysis method and system for nuclear power plant main steam systems. The method includes: constructing a high-dimensional simulation model of the nuclear power plant's main steam system; acquiring simulation model output data after simulating the high-dimensional simulation model with input parameters for each set of model parameters; constructing a dataset based on the corresponding regression relationship between model parameters and simulation model output data; constructing a KAN model; training a sensitivity-converged KAN model based on the dataset, where the input to the KAN model is the model parameters and the output is the surrogate model's predicted output data; performing undirected and directed gradient sensitivity analysis based on the predicted output of the trained sensitivity-converged KAN model to assess the importance of each model parameter; and combining KAN with sensitivity analysis, utilizing the structural interpretability of the KAN model, constructing a parameter sensitivity metric by calculating the gradient information of the model output with respect to the input parameters, thereby achieving a quantitative assessment of the impact of key parameters on the nuclear power plant's main steam system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of nuclear power system analysis, and in particular to a method and system for gradient parameter sensitivity analysis of nuclear power main steam systems based on KAN. Background Technology

[0002] The main steam system is a critical component for the safe and stable operation of a nuclear power plant. Accurately predicting its operating parameters using simulation models is of great significance for monitoring system performance and improving the safety and economy of nuclear power plants. In the nuclear industry, simulation models are typically used to monitor the main steam system's condition to prevent malfunctions from significantly impacting overall safety and reliability. Modeling of the main steam system is usually based on the conservation equations of mass, momentum, and energy to simulate changes in parameters such as steam flow, pressure, and temperature. The modeling process typically involves a large number of model parameters to ensure the accuracy of the simulation model. These parameters are usually derived from equipment characteristics and engineering experience, exhibiting significant high-dimensional and nonlinear characteristics. With continuous unit operation and equipment aging, the model parameters deviate from the actual unit operation, causing discrepancies between the model output and unit data. It is necessary to use unit data and data assimilation methods such as variational methods and ensemble Kalman filtering to calibrate the model parameters to ensure that the model output is consistent with unit operation. However, the high-dimensional and nonlinear characteristics of the model parameters make the calibration process inefficient. Therefore, identifying the key parameters that play a dominant role in the main steam system model output is crucial to improving the efficiency of model parameter calibration, and sensitivity analysis provides an effective method for this process.

[0003] Sensitivity analysis methods are mainly divided into two categories: local sensitivity analysis and global sensitivity analysis. Local sensitivity analysis methods typically quantify parameter importance based on small perturbations or correlation measurements around a reference point, thus reflecting local input-output relationships while ignoring parameter interactions. Global sensitivity analysis methods consider the overall impact of changes in all model parameters on the output, while accurately identifying and ranking key parameters in the model. Local sensitivity analysis is more efficient but has lower accuracy, while global sensitivity analysis requires extensive model evaluation to achieve satisfactory accuracy, which presents a significant computational challenge when model evaluation costs are high. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a KAN-based gradient parameter sensitivity analysis method for nuclear power plant main steam systems. It offers a highly efficient analysis approach that combines global and local perspectives, improving the accuracy and efficiency of sensitivity analysis of nuclear power plant main steam system parameters. While ensuring analytical accuracy, it also considers the sensitivity of computational efficiency. The data-driven proxy model effectively solves the limitations of low accuracy and computational efficiency caused by a large number of calls.

[0005] According to a first aspect of the present invention, a method for sensitivity analysis of gradient parameters in a nuclear power plant main steam system based on KAN is provided, comprising: Step 1: Construct a high-dimensional simulation model of the main steam system of the nuclear power plant, obtain the simulation model output data after inputting each set of model parameters into the high-dimensional simulation model for simulation, and construct a dataset based on the corresponding regression relationship between the gradient data and the simulation model output data; Step 2: Construct a KAN model. Train the KAN model with sensitivity convergence based on the dataset. The input of the KAN model is the model parameters, and the output is the simulation model output data. Step 3: Based on the predicted output of the KAN model that has achieved sensitivity convergence during training, perform undirected and directed gradient sensitivity analysis to evaluate the importance of each model parameter.

[0006] Based on the above technical solution, the present invention can also be improved as follows.

[0007] Optionally, in step 2, during the training of the KAN model, the Adam optimizer and mean squared error loss function are used for training, backpropagation, and optimization by loading mini-batch data. The mathematical expression for the mean squared error loss function is as follows:

[0008] in, Indicates the mean square error loss. The number of training samples, For the first The predicted value for each sample, For the first The true value of each sample.

[0009] Optionally, step 2 further includes: selecting parameters that define the overall structure of the KAN model. The parameter representing the order of the spline function. and a parameter representing the number of spline mesh divisions within each node. Ablation experiments were conducted on the KAN model as an important parameter. The impact of different important parameters on the performance of the KAN model was evaluated based on the root mean square relative error function, and the optimal parameter structure setting of the KAN model was obtained. The mathematical expression for the root mean square relative error function is as follows:

[0010] in, The root mean square relative error, For the first One predicted value, For the first A true value, The total number of samples, This is the mean of the true values.

[0011] Optionally, step 1 further includes: constructing a training set, a test set, and a validation set based on the data in the dataset; Step 2 also includes: Based on the optimal structure of the KAN model, the number of rounds N of the KAN model is determined by selecting the model whose loss on the training set and the validation set remains consistent at a low level after N iterations.

[0012] Optionally, step 2 further includes: The Jaccard index, which represents the degree of overlap between two feature sets, is calculated as follows:

[0013] in, The set of dominant features obtained in each round of training the KAN model. This is the set of dominant features of the KAN model described in the final round; The weighted Spearman coefficients of the KAN model in the current round and the KAN model in the final round are calculated as follows:

[0014] in, and Representing features respectively Weighted sensitivity ranking of the KAN model in the current round and the KAN model in the final round. The number of input features; In the specific implementation process The closer the result is to 1, the more consistent the sorting is. The stability and convergence of the sensitivity results of the KAN model under different training rounds are evaluated based on the Jaccard index and the Spearman coefficient.

[0015] Optionally, the process of performing undirected sensitivity analysis in step 3 includes: Step 301: During the forward propagation of the KAN model, calculate the predicted output of the KAN model. ; Step 302: Calculate each input feature through backpropagation. gradient Calculate any th Input features average gradient for:

[0016] in, It is the sample size. It is the first Predicted values ​​for each sample; Step 303, calculate the first... Input features The normalization sensitivity is = ; in, is the total number of input features, and j represents the index of the input feature.

[0017] Optionally, the process of performing directional sensitivity analysis in step 3 includes: Step 304: Calculate the influence considering both positive and negative directions, specifically the positive influence. negative impact The calculation formulas are as follows:

[0018]

[0019] Step 305: Calculate the absolute value of the negative impact. Then calculate the positive effects separately. and negative impact The normalization results are respectively and ; Step 306: The overall influence of the feature is represented by the difference between positive and negative influences, which is called signature influence. :

[0020] Step 307: Calculate the normalized result of the signature effect. for .

[0021] Optionally, step 3 may be followed by: Step 4, calculate the total Sobol index. The sensitivity results were then ranked, and the total Sobol index was calculated. Similarity to the sensitivity results calculated using the KAN-GSA framework.

[0022] Optionally, step 4 includes: Step 401, calculate the... The first-order Sobol exponent of the parameters Total Sobol Index They are respectively:

[0023]

[0024] in, To remove All other parameters besides; When other parameters are fixed, The resulting output variance; This indicates taking the expectation of other parameters; This represents the high-dimensional simulation model; Step 402, the formula for calculating similarity is:

[0025] in, For parameter index, For the number of parameters, For method The following parameter rankings For method The following parameter rankings For method Sensitivity index under, For method Sensitivity index; similarity results The smaller the value, the better the representation. and methods The more consistent the order of the obtained parameters, the better.

[0026] According to a second aspect of the present invention, a KAN-based gradient parameter sensitivity analysis system for nuclear power main steam systems is provided, comprising: a dataset construction module, a KAN model construction module, and a sensitivity analysis module; The dataset construction module is used to construct a high-dimensional simulation model of the main steam system of a nuclear power plant, obtain the simulation model output data after inputting each set of model parameters into the high-dimensional simulation model for simulation, and construct a dataset based on the corresponding regression relationship between the gradient data and the simulation model output data. The KAN model construction module is used to construct a KAN model, and to train a KAN model with sensitivity convergence based on the dataset. The input of the KAN model is the model parameters, and the output is the simulation model output data. The sensitivity analysis module is used to perform undirected and directed gradient sensitivity analysis based on the predicted output of the KAN model obtained from training and its sensitivity convergence, and to evaluate the importance of each model parameter.

[0027] This invention provides a KAN-based gradient parameter sensitivity analysis method and system for nuclear power main steam systems. By constructing a KAN model and combining it with gradient sensitivity analysis, the method can efficiently perform parameter sensitivity analysis on nuclear power main steam systems.

[0028] The provided KAN-GSA (KAN-based Gradient Sensitivity Analysis) framework, through undirected and directed sensitivity value calculations and comparative experiments, enables sensitivity analysis of complex parameter effects, revealing the contribution of input parameters to the output. By plotting dotted-line graphs to illustrate the magnitude of input changes on the output, it helps engineers quickly identify the magnitude and direction of influence of key parameters, improving the calibration accuracy and efficiency of the main steam system simulation model.

[0029] By combining traditional Sobol sensitivity analysis methods, the total Sobol index is used. The calculations and result ranking validated the accuracy and reliability of the KAN-GSA framework. This method does not rely on a large number of outlier samples for training; the training phase is based solely on normal operating data. Sensitivity analysis is performed using the KAN-GSA method, and the results are compared with those of the Sobol method during the testing phase, ensuring the accuracy of the analysis results.

[0030] The designed sensitivity analysis framework not only handles the high-dimensional, nonlinear, and strongly coupled parameter space of nuclear power main steam systems, but also possesses good adaptability and robustness. The KAN-GSA framework can perform accurate parameter sensitivity analysis under complex operating conditions in nuclear power plants, and through effective sensitivity value calculation, it improves the model's adaptability to different parameter variations and system dynamics.

[0031] The proposed KAN-GSA framework can be rapidly deployed in real-time monitoring and optimization of nuclear power plants, with low computational cost and rapid response capabilities. Through optimization calculation and prediction using the KAN model, this framework is applicable to real-time monitoring and sensitivity analysis of different parameters in nuclear power plant systems, providing rapid support for decision-making and improving the safety and operational stability of nuclear power plant systems. Attached Figure Description

[0032] Figure 1 A flowchart illustrating an embodiment of a KAN-based gradient parameter sensitivity analysis method for nuclear power plant main steam systems provided by this invention; Figure 2 A MultKAN model structure diagram provided in an embodiment of the present invention; Figure 3This is a schematic diagram illustrating the tracking of the training loss and validation loss of the three outputs of the optimally structured KAN model, as provided in an embodiment of the present invention. Figure 4(a) shows the results of the Jaccard index, an indicator for judging sensitivity convergence, provided in the embodiment of the present invention. Figure 4(b) shows the results of the Spearman correlation coefficient, an index for judging sensitivity convergence provided in the embodiment of the present invention. Figure 5 A graph showing the results of sensitivity value calculation without considering direction, provided for an embodiment of the present invention; Figure 6 A graph showing the sensitivity value calculation results considering the direction provided in this embodiment of the invention; Figure 7 The figure shows the comparative experimental results of Sobol provided in the embodiments of the present invention. Detailed Implementation

[0033] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0034] Figure 1 This is a flowchart illustrating an embodiment of a KAN-based gradient parameter sensitivity analysis method for nuclear power plant main steam systems provided by the present invention. Figure 1 This demonstrates the steps of the model initialization and sensitivity analysis modules, combined with... Figure 1 It can be seen that the analytical method includes: Step 1: Construct a high-dimensional simulation model of the main steam system of the nuclear power plant, obtain the simulation model output data after inputting the model parameters of each group into the high-dimensional simulation model for simulation, and construct a dataset based on the corresponding regression relationship between gradient data and simulation model output data. Step 2: Construct the KAN model. Train the KAN model based on the dataset to obtain a sensitivity-converged model. The input of the KAN model is the model parameters, and the output is the simulation model output data. Step 3: Based on the predicted output of the sensitive convergent KAN model obtained from training, perform undirected and directed gradient sensitivity analysis to evaluate the importance of each model parameter.

[0035] This invention provides a KAN-based gradient parameter sensitivity analysis method for nuclear power plant main steam systems. Combining KAN and sensitivity analysis, it leverages the structural interpretability of the KAN model to construct a parameter sensitivity metric by calculating the gradient information of the model output with respect to input parameters. This enables a quantitative assessment of the impact of key parameters on the nuclear power plant main steam system. Compared to traditional global sensitivity analysis methods that require extensive model sampling and computation, this method can directly extract sensitivity information from the trained model, significantly reducing computational costs while maintaining analytical accuracy. This allows for efficient parameter sensitivity analysis in nuclear power plant main steam systems, significantly improving both the model's prediction accuracy and computational efficiency.

[0036] Example 1

[0037] Embodiment 1 provided by this invention is an embodiment of a KAN-based gradient parameter sensitivity analysis method for nuclear power plant main steam systems, combined with... Figure 1 As can be seen, embodiments of this analytical method include: Step 1: Construct a high-dimensional simulation model of the main steam system of the nuclear power plant, obtain the simulation model output data after inputting the model parameters of each group into the high-dimensional simulation model for simulation, and construct a dataset based on the corresponding regression relationship between gradient data and simulation model output data. In the specific implementation process, 94 model parameters were selected as inputs, and a regression relationship between inputs and outputs was established. A convergent KAN model was trained using simulation data to predict outputs and perform sensitivity analysis.

[0038] In one possible implementation, after the dataset is constructed, a training set, a test set, and a validation set are also constructed based on the data in the dataset; In practice, the data can be divided into a training set and a test set in an 8:2 ratio. The training set and the test set are then standardized. A portion of the training set is further divided into a validation set for model validation during the training process.

[0039] Step 2: Construct the KAN model. Train the KAN model based on the dataset to obtain a sensitivity-converged model. The input of the KAN model is the model parameters, and the output is the simulation model output data. like Figure 2 The diagram shown is a structural diagram of a MultKAN model provided in an embodiment of the present invention. The MultKAN model is an enhanced version of KAN. Its core improvement is the introduction of additional multiplication nodes. It consists of three parts: an input layer, a hidden layer, and an output layer. The input and output layers are consistent with traditional neural networks. MultKAN has the same structure and concept as KAN. The MultKAN model used in this application is collectively referred to as the KAN model.

[0040] In one possible implementation, during the training of the KAN model in step 2, training, backpropagation, and optimization are performed using the Adam optimizer and mean squared error loss function by loading mini-batch data. In practice, the initial values ​​of the number of layers, nodes, and other hyperparameters of the KAN model can be set according to basic recommendations.

[0041] The mathematical expression for the mean squared error loss function is:

[0042] in, Indicates the mean square error loss. The number of training samples, For the first The predicted value for each sample, For the first The true value of each sample.

[0043] In one possible embodiment, step 2 further includes: selecting parameters that define the overall structure of the KAN model. The parameter representing the order of the spline function. and a parameter representing the number of spline mesh divisions within each node. Ablation experiments were conducted on the KAN model as an important parameter. The impact of different important parameters on the performance of the KAN model was evaluated based on the root mean square relative error function, and the optimal parameter structure setting of the KAN model was obtained. The computational efficiency and training performance of the KAN model were balanced.

[0044] The mathematical expression for the root mean square relative error function is:

[0045] in, The root mean square relative error, For the first One predicted value, For the first A true value, The total number of samples, This is the mean of the true values.

[0046] Tables 1 and 2 below show the ablation experiment results provided in this application.

[0047] Table 1. Results of ablation experiments under network structures with different numbers of hidden layers.

[0048] Table 2. Results of ablation experiments under network structures with different values ​​of parameters k and grid.

[0049] Tables 1 and 2 show the index results under various network structure settings and the final selected network structure settings (2 hidden layers, each containing 16 ordinary nodes, and...). ).

[0050] like Figure 3 The diagram illustrates the tracking of training and validation losses for the three outputs of the optimally structured KAN model according to an embodiment of the present invention. Tracking the training and validation losses of the three outputs of the KAN model under the optimal structure determined by ablation experiments shows that the model has good generalization performance and no obvious overfitting or underfitting. Therefore, a convergent model trained for 50 epochs is subsequently adopted.

[0051] In one possible embodiment, step 2 further includes: Based on the optimal KAN model structure, the number of iterations N is determined by selecting models where the losses on the training and validation sets remain consistent at a low level after N iterations. This ensures that the KAN model has good generalization performance and does not exhibit significant overfitting or underfitting.

[0052] In one possible embodiment, step 2 further includes: The Jaccard index, which represents the degree of overlap between two feature sets, is calculated as follows:

[0053] in, The set of dominant features obtained in each round of training the KAN model. This is the set of dominant features for the final round of the KAN model; The weighted Spearman coefficients of the KAN model at the current round number and the KAN model at the final round number are calculated as follows:

[0054] in, and Representing features respectively Weighted sensitivity ranking under the current round number KAN model and the final round number KAN model. The number of input features; In the specific implementation process The closer the result is to 1, the more consistent the sorting is. The stability and convergence of the sensitivity results of the KAN model under different training rounds were evaluated based on the Jaccard index and Spearman coefficient.

[0055] To quantitatively evaluate the stability and convergence of sensitivity results across different training rounds, two complementary indices, the Jaccard index and the weighted Spearman correlation coefficient, were selected. The Jaccard index was used to measure the overlap between two feature sets, and an exponential weight was added to the sensitivity to enhance the influence of highly sensitive features. The weighted sensitivity was defined as follows: To ensure the stable convergence and reliable results of the sensitivity analysis, the Jaccard index and Spearman correlation coefficient are introduced as indicators to judge the convergence of sensitivity.

[0056] Figures 4(a) and 4(b) show the results of the Jaccard index and Spearman correlation coefficient, which are indicators for judging sensitivity convergence provided in the embodiments of the present invention. Figures 4(a) and 4(b) show the sensitivity convergence index results of the three outputs after training the model for 50 rounds, indicating that the parameter sensitivity structure learned by the KAN model has reached convergence at both the dominant parameter set and overall ranking levels.

[0057] Step 3: Based on the predicted output of the sensitive convergent KAN model obtained from training, perform undirected and directed gradient sensitivity analysis to evaluate the importance of each model parameter.

[0058] In one possible embodiment, step 3, which involves performing undirected sensitivity analysis, includes: Step 301: During the forward propagation of the KAN model, calculate the predicted output of the KAN model. ; Step 302: Calculate each input feature through backpropagation. gradient This gradient represents the model's predicted output. Relative to input features The rate of change. The absolute value of the gradient for each feature. Calculate any th Input features average gradient for:

[0059] in, It is the sample size. It is the first Predicted values ​​for each sample; Step 303, calculate the first... Input features The normalization sensitivity is = ; in, is the total number of input features, and j represents the index of the input feature. It is the first The average absolute gradient of each feature.

[0060] like Figure 5 The figure shown is a graph illustrating the sensitivity value calculation results without considering direction, provided by an embodiment of the present invention. Figure 5 The results show that different outputs are dominated by a few core parameters, while the contributions of other parameters to output changes are relatively limited. At the same time, the sets of highly sensitive parameters corresponding to each output differ significantly, reflecting that different output quantities in the main steam system are controlled by different physical processes.

[0061] The process of conducting directional sensitivity analysis includes: Step 304: Calculate the influence considering both positive and negative directions, specifically the positive influence. negative impact The calculation formulas are as follows:

[0062]

[0063] Step 305: Calculate the absolute value of the negative impact. Then calculate the positive effects separately. and negative impact The normalization results are respectively and ; Step 306: The overall influence of the feature is represented by the difference between positive and negative influences, which is called signature influence. :

[0064] Step 307: Calculate the normalized result of the signature effect. for .

[0065] By calculating undirected and directed sensitivity values, rapid and accurate analysis results can be provided for complex nuclear power main steam systems. The KAN-GSA framework is applied for parameter gradient sensitivity analysis, introducing undirected and directed indices for gradient sensitivity. By calculating these indices, the contribution of each input parameter to the system output is quantified, and the importance of each parameter is assessed, thereby improving the accuracy and computational efficiency of the main steam system simulation model.

[0066] like Figure 6 The figure shown is a graph illustrating the sensitivity value calculation results considering the direction provided in an embodiment of the present invention. Figure 6 This study demonstrates how the model parameters affect the output results without considering the directional results, thus providing a more comprehensive basis for the interpretation and parameter control of the main steam system simulation model.

[0067] In one possible embodiment, step 3 is followed by: Step 4, calculate the total Sobol index. The sensitivity results were then ranked, and the total Sobol index was calculated. Similarity to the sensitivity results calculated using the KAN-GSA framework.

[0068] By introducing the traditional Sobol sensitivity analysis method, the similarity between the results of the two methods was calculated, and the accuracy and effectiveness of the results of the KAN-GSA framework were verified, providing an effective solution for the optimized operation and parameter control of the main steam system of nuclear power plants.

[0069] In one possible embodiment, step 4 includes: Step 401, calculate the... The first-order Sobol exponent of the parameters Total Sobol Index They are respectively:

[0070]

[0071] in, To remove All other parameters besides; When other parameters are fixed, The resulting output variance; This indicates taking the expectation of other parameters; Represents a high-dimensional simulation model; In the specific implementation process, the total Sobol index is used to measure the total impact of input features on the output of the high-dimensional simulation model (including its direct impact and interaction with other features). Subsequent experiments use the total Sobol index as an indicator to compare with the results of the KAN-GSA method.

[0072] Step 402, the formula for calculating similarity is:

[0073] in, For parameter index, For the number of parameters, For method The following parameter rankings For method The following parameter rankings For method Sensitivity index under, For method Sensitivity index; similarity results The smaller the value, the better the representation. and methods The more consistent the order of the obtained parameters, the better.

[0074] By introducing the traditional Sobol sensitivity analysis method and calculating the ST index, this method can compare and verify the results of the KAN-GSA framework, ensuring the accuracy and reliability of the analysis.

[0075] like Figure 7 The figure shown is a comparison experiment result diagram of Sobol provided by the embodiment of the present invention. Figure 7 The study demonstrates that the influence of key model parameters in the two methods highly overlaps, and the similarity of all model parameters is approximately 0.05. This verifies the stability and consistency of the proposed KAN-GSA method in parameter importance identification, indicating that the data-driven surrogate model can balance the advantages of computational efficiency and analytical accuracy.

[0076] Example 2

[0077] Embodiment 2 of the present invention is an embodiment of a KAN-based gradient parameter sensitivity analysis system for nuclear power main steam systems. The embodiment of the system includes: a dataset construction module, a KAN model construction module, and a sensitivity analysis module. The dataset construction module is used to build a high-dimensional simulation model of the main steam system of a nuclear power plant, obtain the simulation model output data after inputting the model parameters of each group into the high-dimensional simulation model for simulation, and construct the dataset based on the corresponding regression relationship between gradient data and simulation model output data. The KAN model building module is used to build KAN models. It trains KAN models with sensitivity convergence based on datasets. The input of the KAN model is the model parameters, and the output is the simulation model output data. The sensitivity analysis module is used to perform undirected and directed analysis of gradient sensitivity based on the predicted output of the KAN model that has converged in sensitivity obtained from training, and to evaluate the importance of each model parameter.

[0078] It is understood that the KAN-based gradient parameter sensitivity analysis system for nuclear power main steam systems provided by this invention corresponds to the KAN-based gradient parameter sensitivity analysis methods for nuclear power main steam systems provided in the foregoing embodiments. The relevant technical features of the KAN-based gradient parameter sensitivity analysis system for nuclear power main steam systems can be referred to the relevant technical features of the KAN-based gradient parameter sensitivity analysis methods for nuclear power main steam systems, and will not be repeated here.

[0079] This invention provides a KAN-based gradient parameter sensitivity analysis method and system for nuclear power plant main steam systems. The method proceduralizes the parameter sensitivity analysis process of nuclear power plant main steam systems through steps such as data processing, model training and storage, parameter sensitivity analysis, and result verification. This ensures that the impact of each parameter on system performance can be accurately assessed, thereby optimizing the system's operating parameters and improving the system's performance and stability.

[0080] In the data processing phase, the simulation data was first divided into training and test sets in an 8:2 ratio to ensure the independence of model training and performance evaluation. Subsequently, both the training and test sets were standardized to eliminate the influence of unit dimensions and improve the stability of model training. Simultaneously, a portion of the training set was further partitioned as a validation set for monitoring model performance and evaluating generalization ability during training. Training was performed using mini-batch data loading to improve computational efficiency and enhance the stability of parameter updates. This phase ensured that the data input was standardized and reasonably distributed, laying a reliable foundation for model training and subsequent sensitivity analysis. In the model building phase, a model was built based on KAN, and the initial number of network layers, nodes, and other key hyperparameters were set according to empirical recommendations. The model was trained using the Adam optimizer and the mean squared error loss function.

[0081] Ablation experiments were conducted around three key structural parameters to systematically compare the impact of different structural combinations on the model's predictive performance. The root mean square relative error (RMSE) was used as an evaluation metric to select the optimal network structure. Under the optimal structure setting, the changing trends of the loss curves on the training and validation sets were further observed. The best training epochs were selected when both training and validation losses were at low levels and showed consistent trends, avoiding significant overfitting or underfitting. Finally, models with good generalization performance were saved for subsequent sensitivity analysis and comparative experiments. To evaluate the stability and convergence of sensitivity results across different training epochs, two complementary metrics were introduced for quantitative analysis. First, the Jaccard index was used to measure the overlap of the dominant feature sets across different training epochs, assessing the stability of the key parameter identification results. Second, the Spearman correlation coefficient was used to measure the consistency between the sensitivity ranking vectors across different training epochs, and the influence of highly sensitive features was enhanced through exponential weighting, focusing on the stability trend of the core parameter ranking.

[0082] After obtaining the predicted output through forward propagation of the model, the gradient of the output relative to each input feature is calculated using automatic differentiation. The average of the absolute values ​​of the gradients is used to measure the influence of each parameter on the model output, and normalization is performed to obtain a relative sensitivity ranking. Based on this, directionality analysis is introduced, decomposing the gradient into positive and negative effects. The average contribution of each effect is calculated and normalized, further constructing a signature influence index to characterize the positive or negative regulatory effect of parameters on system performance. This step not only quantifies the intensity of parameter influence but also reveals its direction, thus providing a more comprehensive understanding of the system's internal mechanisms.

[0083] Simultaneously, the Sobol sensitivity analysis method was introduced to calculate the overall sensitivity index of each parameter, which measures the overall impact of the parameters on the model output. Subsequent experiments used the overall Sobol index as a control index to systematically compare with the KAN gradient sensitivity results to verify the consistency and reliability of the method. Furthermore, to compare the ranking consistency between the KAN-GSA method and the Sobol method, a normalized similarity index was introduced to quantitatively analyze the degree of difference between the parameter rankings obtained by the two methods, thereby verifying the effectiveness of gradient sensitivity analysis at the global level.

[0084] This application discloses a KAN-based gradient parameter sensitivity analysis method for nuclear power plant main steam systems, applicable to complex industrial systems such as nuclear power plants. Specifically, a KAN-GSA sensitivity analysis framework is proposed to improve the accuracy of analysis and the computational efficiency of the model under complex operating conditions. This method, through a KAN surrogate model, enables efficient and accurate parameter sensitivity analysis in the main steam system, significantly improving the prediction accuracy and computational efficiency of the system model. The present invention provides a KAN-based gradient parameter sensitivity analysis method and system for nuclear power plant main steam systems, which, by constructing a KAN model combined with gradient sensitivity analysis, can efficiently perform parameter sensitivity analysis on nuclear power plant main steam systems.

[0085] The provided KAN-GSA (KAN-based Gradient Sensitivity Analysis) framework, through undirected and directed sensitivity value calculations and comparative experiments, enables sensitivity analysis of complex parameter effects, revealing the contribution of input parameters to the output. By plotting dotted-line graphs to illustrate the magnitude of input changes on the output, it helps engineers quickly identify the magnitude and direction of influence of key parameters, improving the calibration accuracy and efficiency of the main steam system simulation model.

[0086] By combining traditional Sobol sensitivity analysis methods, the total Sobol index is used. The calculations and result ranking validated the accuracy and reliability of the KAN-GSA framework. This method does not rely on a large number of outlier samples for training; the training phase is based solely on normal operating data. Sensitivity analysis is performed using the KAN-GSA method, and the results are compared with those of the Sobol method during the testing phase, ensuring the accuracy of the analysis results.

[0087] The designed sensitivity analysis framework not only handles the high-dimensional, nonlinear, and strongly coupled parameter space of nuclear power main steam systems, but also possesses good adaptability and robustness. The KAN-GSA framework can perform accurate parameter sensitivity analysis under complex operating conditions in nuclear power plants, and through effective sensitivity value calculation, it improves the model's adaptability to different parameter variations and system dynamics.

[0088] The proposed KAN-GSA framework can be rapidly deployed in real-time monitoring and optimization of nuclear power plants, with low computational cost and rapid response capabilities. Through optimization calculation and prediction using the KAN model, this framework is applicable to real-time monitoring and sensitivity analysis of different parameters in nuclear power plant systems, providing rapid support for decision-making and improving the safety and operational stability of nuclear power plant systems.

[0089] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0090] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0092] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0093] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0094] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0095] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for sensitivity analysis of gradient parameters in a nuclear power plant main steam system based on KAN, characterized in that, The analytical method includes: Step 1: Construct a high-dimensional simulation model of the main steam system of the nuclear power plant, obtain the simulation model output data after inputting each set of model parameters into the high-dimensional simulation model for simulation, and construct a dataset based on the corresponding regression relationship between the gradient data and the simulation model output data; Step 2: Construct a KAN model. Train the KAN model with sensitivity convergence based on the dataset. The input of the KAN model is the model parameters, and the output is the simulation model output data. Step 3: Based on the predicted output of the KAN model that has achieved sensitivity convergence during training, perform undirected and directed gradient sensitivity analysis to evaluate the importance of each model parameter.

2. The analytical method according to claim 1, characterized in that, In step 2, during the training of the KAN model, the Adam optimizer and mean squared error loss function are used for training, backpropagation, and optimization by loading mini-batch data. The mathematical expression for the mean squared error loss function is as follows: in, This represents the mean squared error loss. The number of training samples, For the first The predicted value for each sample, For the first The true value of each sample.

3. The analytical method according to claim 1, characterized in that, Step 2 also includes: selecting parameters that define the overall structure of the KAN model. The parameter representing the order of the spline function. and a parameter representing the number of spline mesh divisions within each node. Ablation experiments were conducted on the KAN model as an important parameter. The impact of different important parameters on the performance of the KAN model was evaluated based on the root mean square relative error function, and the optimal parameter structure setting of the KAN model was obtained. The mathematical expression for the root mean square relative error function is as follows: in, The root mean square relative error, For the first One predicted value, For the first A true value, The total number of samples, This is the mean of the true values.

4. The analytical method according to claim 3, characterized in that, Step 1 further includes: constructing a training set, a test set, and a validation set based on the data in the dataset; Step 2 also includes: Based on the optimal structure of the KAN model, the number of rounds N of the KAN model is determined by selecting the model whose loss on the training set and the validation set remains consistent at a low level after N iterations.

5. The analytical method according to claim 1, characterized in that, Step 2 also includes: The Jaccard index, which represents the degree of overlap between two feature sets, is calculated as follows: in, The set of dominant features obtained in each round of training the KAN model. This is the set of dominant features of the KAN model described in the final round; The weighted Spearman coefficients of the KAN model in the current round and the KAN model in the final round are calculated as follows: in, and Representing features respectively Weighted sensitivity ranking of the KAN model in the current round and the KAN model in the final round. The number of input features; The stability and convergence of the sensitivity results of the KAN model under different training rounds are evaluated based on the Jaccard index and the Spearman coefficient.

6. The analytical method according to claim 1, characterized in that, The process of performing undirected sensitivity analysis in step 3 includes: Step 301: During the forward propagation of the KAN model, calculate the predicted output of the KAN model. ; Step 302: Calculate each input feature through backpropagation. gradient Calculate any th Input features average gradient for: in, It is the sample size. It is the first Predicted values ​​for each sample; Step 303, calculate the first... Input features The normalization sensitivity is = ; in, is the total number of input features, and j represents the index of the input feature.

7. The analytical method according to claim 6, characterized in that, The process of performing directional sensitivity analysis in step 3 includes: Step 304: Calculate the influence considering both positive and negative directions, specifically the positive influence. negative impact The calculation formulas are as follows: Step 305: Calculate the absolute value of the negative impact. Then calculate the positive effects separately. and negative impact The normalization results are respectively and ; Step 306: The overall influence of the feature is represented by the difference between positive and negative influences, which is called signature influence. : Step 307: Calculate the normalized result of the signature effect. for .

8. The analytical method according to claim 1, characterized in that, Step 3 is followed by: Step 4, calculate the total Sobol index. The sensitivity results were then ranked, and the total Sobol index was calculated. Similarity to the sensitivity results calculated by the KAN-GSA method.

9. The analytical method according to claim 8, characterized in that, Step 4 includes: Step 401, calculate the... The first-order Sobol exponent of the parameters Total Sobol Index They are respectively: in, To remove All other parameters besides; When other parameters are fixed, The resulting output variance; This indicates taking the expectation of other parameters; This represents the high-dimensional simulation model; Step 402, the formula for calculating similarity is: in, For parameter index, For the number of parameters, For method The following parameter rankings For method The following parameter rankings For method Sensitivity index under, For method Sensitivity index; similarity results The smaller the value, the better the representation. and methods The more consistent the order of the obtained parameters, the better.

10. A gradient parameter sensitivity analysis system for nuclear power plant main steam systems based on KAN, characterized in that, The analysis system includes: a dataset construction module, a KAN model construction module, and a sensitivity analysis module; The dataset construction module is used to construct a high-dimensional simulation model of the main steam system of a nuclear power plant, obtain the simulation model output data after inputting each set of model parameters into the high-dimensional simulation model for simulation, and construct a dataset based on the corresponding regression relationship between the gradient data and the simulation model output data. The KAN model construction module is used to construct a KAN model, and to train a KAN model with sensitivity convergence based on the dataset. The input of the KAN model is the model parameters, and the output is the simulation model output data. The sensitivity analysis module is used to perform undirected and directed gradient sensitivity analysis based on the predicted output of the KAN model obtained from training and its sensitivity convergence, and to evaluate the importance of each model parameter.